Vehicle counting based on a stereo vision depth maps for parking management

Abstract

Automated parking management systems provide convenience and efficiency, and such systems are increasingly being deployed in modern urban areas. To facilitate the crucial function of vehicle counting in several applications, we developed a novel mechanism for counting vehicles based on stereoscopic computer vision with depth perception. In this study, we first established depth maps of pairs of images captured using stereo cameras through a scene flow-based approach. Next, we designed a modified sigmoid function to change the histogram distribution in the obtained depth maps by using the disparity threshold estimated from a disparity calibration board. Then, we proposed a vehicle counting mechanism using the modified disparity histogram; this mechanism can be used to easily determine the presence of a vehicle. Consequently, we applied the proposed vehicle detection and counting method to a surveillance camera and used it to determine whether vehicles were approaching an entrance; this camera captured a clear photograph of each license plate, which was then used for automatic recognition. The proposed system was evaluated using nine sets of video data recorded in an indoor parking garage and an outdoor parking lot. The experimental results quantified our method’s high performance and robustness in vehicle counting. For the indoor parking garage, the precision and recall were 99.56% and 98.29%, respectively. For the outdoor parking lot environment, the vehicle counting precision and recall were 98.85% and 98.85%, respectively. Our method was able to avoid counting errors when distinguishing between closely spaced adjacent vehicles.